Fig. 2. Examples of chemical insights extracted by SpookyNet.

a Visualization of the learned local chemical potential for ethanol (see methods). The individual contributions of s-, p-, and d-orbital-like interactions are shown (red: low energy, blue: high energy). b Potential energy surface scans obtained by moving an Au₂ dimer over an (Al-doped) MgO surface in different (upright/parallel) configurations (the positions of Mg and O atoms are shown for orientation). SpookyNet learns to distinguish between local and nonlocal contributions to the potential energy, allowing it to model changes of the potential energy surface when the crystal is doped with Al atoms far from the surface. c A model trained on small organic molecules learns general chemical principles that can be transferred to much larger structures outside the chemical space covered by the training data. Here, optimized geometries obtained from SpookyNet trained on the QM7-X database (opaque in color) are shown and compared with reference geometries obtained from ab initio calculations (transparent in gray). As indicated by the low root mean square deviations (RMSD), geometries predicted by SpookyNet are very similar to the reference.